Data is scaling exponentially in fields ranging from genomics to neuroscience to economics. A central question is whether modern machine learning methods can be applied to construct predictive models based on large data sets drawn from complex, natural systems like cells and brains. In machine learning, the predictive power or generalizability of a model is determined by the statistics of training data. In this paper, we ask how predictive inference is impacted when training data is generated by the statistical behavior of a physical system. We develop an information-theoretic analysis of a canonical problem, spin network inference. Our analysis reveals the essential role that thermal fluctuations play in determining the efficiency of predictive inference. Thermal noise drives a system to explore a range of configurations providing `raw’ information for a learning algorithm to construct a predictive model. Conversely, thermal energy degrades information by blurring energetic differences between network states. In general, spin networks have an intrinsic optimal temperature at which inference becomes maximally efficient. Simple active learning protocols allow optimization of network temperature, without prior knowledge, to dramatically increase the efficiency of inference. Our results reveal a fundamental link between physics and information and show how the physical environment can be tuned to optimize the efficiency of machine learning. Physically optimizing inference